fix(openai): handle None arguments in legacy function_call streaming merge

[roli-lpci]

What does this PR do?

_extract_streamed_openai_response (langfuse/openai.py) merges the deprecated function_call
streaming delta with a bare +=:

curr["arguments"] += getattr(tool_call_chunk, "arguments", "")

getattr(obj, "arguments", "") only returns the "" default when the attribute is absent. Some
OpenAI-compatible proxies send a first delta with function_call.arguments explicitly set to
None (not missing) — a real payload shape, not a hypothetical one. In that case
curr["arguments"] is None, and None += str raises TypeError, killing the merge for the
whole streamed completion. The TypeError is currently swallowed by a bare except Exception: pass further up the call stack, so today this fails silently: the generation ends with no output
and no error a caller would notice.

This is the same crash class PR #1339 already fixed for the newer tool_calls array — but that
fix only touched the tool_calls branch a few lines below; the sibling legacy function_call
branch was never patched. The tool_calls branch already guards the accumulator side and skips
None increments:

if tool_arguments is not None:
    function_call["arguments"] = (function_call.get("arguments") or "") + tool_arguments

The fix brings the same protection to the legacy branch, adopting that branch's
(... or "") + ... accumulator guard and coalescing the incoming delta inline so a None (or
missing) arguments on either side can't break the concatenation:

curr["arguments"] = (curr.get("arguments") or "") + (
    getattr(tool_call_chunk, "arguments", None) or ""
)

This guards both sides: the accumulator being None, and the incoming delta's arguments being
None or missing. When both sides are normal strings it is equivalent to the original += — no
behavior change on the happy path. Two-line functional change to a single merge branch; does not
touch the tool_calls branch or any other streaming path.

No existing issue or open PR describes this legacy-function_call-branch crash (see the search
under Verification below); happy to open a tracking issue first if that's preferred for bug-fix PRs.

Type of change

• Bug fix
• New feature
• Breaking change
• Refactor
• Documentation update
• Tooling, CI, or repo maintenance

Verification

List the main commands you ran:

# RED: fix reverted (source only, test kept)
python -m pytest tests/unit/test_openai.py::test_streaming_chat_completion_handles_legacy_function_call_none_arguments -q
# -> 1 failed: TypeError: unsupported operand type(s) for +=: 'NoneType' and 'str' in the legacy
#    function_call merge branch of _extract_streamed_openai_response

# GREEN: fix restored
python -m pytest tests/unit/test_openai.py::test_streaming_chat_completion_handles_legacy_function_call_none_arguments -q
# -> 1 passed

# Full mapped file, regression check
python -m pytest tests/unit/test_openai.py
# -> 35 passed (34 pre-existing + 1 new), no regressions

ruff check langfuse/openai.py tests/unit/test_openai.py
# -> All checks passed!

gh pr list --repo langfuse/langfuse-python --state open --search "function_call None arguments"
# -> empty; #1339 (merged) is the sibling tool_calls fix this mirrors; no PR touches the legacy
#    function_call branch.

Test added: test_streaming_chat_completion_handles_legacy_function_call_none_arguments, mirroring
the existing test_streaming_chat_completion_preserves_tool_calls_after_content pattern but for
the legacy branch, with a first delta carrying function_call.arguments=None.

Checklist

• I self-reviewed the diff using code_review.md.
• I added or updated tests for behavior changes.
• I updated docs, examples, or .env.template if needed. (No public API or documented
  behavior changed; two-line internal guard.)
• I did not hand-edit generated files; if generated files changed, I used the upstream
  regeneration path.
• I did not commit secrets or credentials.

Greptile Summary

This PR fixes a silent TypeError crash in the deprecated function_call streaming-merge branch of _extract_streamed_openai_response when an OpenAI-compatible proxy sends a delta with arguments explicitly set to None. The fix mirrors the identical guard already applied to the tool_calls branch in PR #1339 that was never extended to this legacy sibling branch.

• langfuse/openai.py: Replaces the bare curr["arguments"] += getattr(...) accumulation with a null-safe pattern (curr.get("arguments") or "") + (getattr(...) or ""), protecting both the accumulator and the incoming delta from being None.
• tests/unit/test_openai.py: Adds a regression test that replays the two-delta payload shape (first delta with arguments=None, second with the real JSON string) and asserts correct final output.

Confidence Score: 5/5

Safe to merge — the change is a two-line targeted guard that cannot regress the happy path (string + string produces the same result as before), and the new test directly validates the failing case.

The fix is a minimal null-coalescing guard that mirrors an already-proven pattern from the tool_calls branch. The initialization path at line 807 could still store None for arguments on a single-delta stream, but that pre-existing edge produces None output rather than a crash and is out of scope for this PR. No other correctness issues found.

No files require special attention.

Sequence Diagram

%%{init: {'theme': 'neutral'}}%%
sequenceDiagram
    participant Proxy as OpenAI-compatible Proxy
    participant Extract as _extract_streamed_openai_response
    participant Accum as function_call accumulator

    Proxy->>Extract: "chunk 1: function_call.arguments = None"
    Extract->>Accum: "if not curr → initialize {name, arguments: None}"

    Proxy->>Extract: "chunk 2: function_call.arguments = '{"city":"Berlin"}'"
    Note over Extract,Accum: BEFORE fix: None += str → TypeError (swallowed silently)
    Note over Extract,Accum: AFTER fix: (None or "") + (str or "") → '{"city":"Berlin"}'
    Extract->>Accum: "curr["arguments"] = "" + '{"city":"Berlin"}'"
    Accum-->>Extract: "{name: "get_weather", arguments: '{"city":"Berlin"}'}"

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
sequenceDiagram
    participant Proxy as OpenAI-compatible Proxy
    participant Extract as _extract_streamed_openai_response
    participant Accum as function_call accumulator

    Proxy->>Extract: "chunk 1: function_call.arguments = None"
    Extract->>Accum: "if not curr → initialize {name, arguments: None}"

    Proxy->>Extract: "chunk 2: function_call.arguments = '{"city":"Berlin"}'"
    Note over Extract,Accum: BEFORE fix: None += str → TypeError (swallowed silently)
    Note over Extract,Accum: AFTER fix: (None or "") + (str or "") → '{"city":"Berlin"}'
    Extract->>Accum: "curr["arguments"] = "" + '{"city":"Berlin"}'"
    Accum-->>Extract: "{name: "get_weather", arguments: '{"city":"Berlin"}'}"

Loading

_{Reviews (1): Last reviewed commit: "test(openai): rebase legacy function_cal..." | Re-trigger Greptile}

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}